From 0dc7f258a35628ad61522f6bb5a40468d8629e6e Mon Sep 17 00:00:00 2001 From: Hyperling Date: Wed, 14 Oct 2026 05:34:41 -0700 Subject: [PATCH] llm, reduce and simplify concurrecy and parallel as to not waste system resources or time. --- tasks/workstation/linux/software/llm.yml | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/tasks/workstation/linux/software/llm.yml b/tasks/workstation/linux/software/llm.yml index 55c6747..8a6816f 100644 --- a/tasks/workstation/linux/software/llm.yml +++ b/tasks/workstation/linux/software/llm.yml @@ -53,10 +53,14 @@ dummy_website: 'https://hyperling.com' # Good to keep low to prevent swap usage. Going to have a minimum of 3 # though, one for Cline, two for Continue (Agent + Autocomplete). - llm_concurrency: "{{ [(ansible_memtotal_mb / 32000) | round(0, method='ceil'), 3] | max | int }}" + # 20261014: Going down to 2. Either 1 strong and 1 weak (autocomplete), + # or 1 coder and 1 general. No need to exceed this even on good HW. + llm_concurrency: 2 #"{{ [(ansible_memtotal_mb / 32000) | round(0, method='ceil'), 3] | max | int }}" # Allow having multiple chats open without them just waiting each other. # Also setting this to a minimum of 3 for using Codium with extensions. - llm_parallel: "{{ [(ansible_processor_vcpus / 2) | round(0, method='ceil'), 3] | max | int }}" + # 20261014: Going down to 1 since only using CPUs, no GPUs. Otherwise + # the models seem to take excessively long to answer even small things. + llm_parallel: 1 #"{{ [(ansible_processor_vcpus / 2) | round(0, method='ceil'), 3] | max | int }}" # Most LLMs take a coefficient of about 1.2 to 1.5 of their file size in # RAM, which is usually less than guessing a usage of weight*1000. # Example: